Cloud Development Environment Platform  ·  View 11 of 21  ·  Data

Storage Zones by Ownership and Rebuildability

Three bands: what cannot be lost, what costs time to lose, and what is on its way to deletion.

Editable source SVG draw.io All views
Cannot be lost · the workspace itself Workspace content Filesystem snapshots RPO 5 min · 180 TB/d Memory images best effort Control-plane truth Workspace registry RPO 0 · replicated Tenants, seats, policy RPO 0 Evidence Audit records immutable · 5 y Rebuildable · losing it costs time, not truth Start-time caches Prebuild artefacts rebuild ≤ 4 h Base images digest-pinned Proxy caches Registry pull-through recomputed Operational Telemetry raw 14 d Session records 400 d Lifecycle · archived, then gone Warm Not resumed ≤ 30 d $0.004/h Cold Archived to 90 d rehydrate on ask Deleted Deleted at 90 d unless extended tiering retention dedup base Storage Zones by Ownership and Rebuildability Data store Risk / gap batch synchronous Everything in the middle band may be deleted on a bad day; the top band is what the platform exists to keep. v 1.0 · owner Platform Architecture · date 2026-10

Decisions

  • Zones are drawn by rebuildability rather than by technology, because that is the question a backup budget actually answers: four of these stores can be lost and recomputed, and two cannot.
  • Memory images are in the top band but marked best effort: losing one costs a cold boot, not a workspace, which is what makes the decay window in ADR-03 affordable.
  • Archival is a zone, not a job: warm to 30 days, cold to 90, then deleted unless the tenant extends it.

Assumptions

  • 2,400,000 hibernated workspaces retained at ≤ $0.004 per workspace-hour in storage and metadata.
  • Snapshot durability ≥ 11 nines with integrity verified on every write and restore; prebuild capacity for the top 100 repositories rebuildable in 4 hours.

Risks

  • Snapshot storage for workspaces that are never resumed is pure waste; the target is ≤ 3% of storage spend, and it needs watching rather than assuming.
  • Customer-managed key revocation makes a tenant's whole top band unreadable by design — correct, and a support incident waiting to be mistaken for data loss.